无线电工程

火灾引发的经济与人员损伤始终是社会的棘手问题，迫切需要能够实时、准确监控火灾发生的方案。针对城市火灾场景复杂、目标小和定位要求高等问题，提出了一种改进YOLOv5的城市火灾场景下烟火目标检测算法。整理收集到的网络数据、构建数据集，并进行数据增强。基于YOLOv5s算法模型，重构网络结构，增加小目标检测层，使模型更加关注小目标的检测。嵌入了压缩与激励网络(Squeeze-and-Excitation Network, SENet),使YOLOv5模型的检测精度进一步提升。讨论了SENet添加位置的问题。实验结果表明，改进YOLOv5算法的精确率达到了93.7%,与原YOLOv5s相比召回率和平均精确度分别提高了1.9%、1.6%;在网络中添加注意力模块的位置不同，所产生的模型效果也不同。

The economic losses and personnel injuries caused by fires have always been a thorny issue in society, and there is an urgent need for solutions that can monitor the occurrence of fires in real-time and accurately. A modified YOLOv5 algorithm for detecting pyrotechnic targets in urban fire scenes is proposed to address the problems of complexity, small targets, and high positioning requirements in urban fire scenes. Firstly, the collected network data is organized, the dataset is constructed, and the data enhancement is performed. Then, based on the YOLOv5s algorithm model, the network structure is reconstructed and a small object detection layer is added to make the model pay more attention to small object detection. Finally, the Squeeze-and-Excitation Network(SENet) are embedded to further improve the detection accuracy of the YOLOv5 model. In addition, the issue of adding locations for SENet is also discussed. The experimental results show that the accuracy of the improved YOLOv5 algorithm has reached 93.7%, and compared with the original YOLOv5s, the recall rate and average accuracy have increased by 1.9% and 1.6%; it is found that adding the attention module at different locations in the network produces different modeling effects.